gpu::matchTemplate returns inconsistent result with cv::matchTemplate (Bug #1341)

I found two sets of images that cause gnu::matchTemplate and cv::matchTemplate return different results using CV_TM_CCOEFF_NORMED. The GPU results are even not normalized correctly.

gpu::matchTemplate(gSource,gTarget,gResult,CV_TM_CCOEFF_NORMED);
   gpu::minMaxLoc(gResult, &min, &max, &min_loc, &max_loc);
   cout << "gpu min: " << min << ", max: " << max << endl;
   cout << "gpu min location: " << min_loc.x << "," << min_loc.y
           << ", max location: "  << max_loc.x << "," << max_loc.y  << endl;

matchTemplate(source,target,result,CV_TM_CCOEFF_NORMED);
   minMaxLoc(result, &min, &max, &min_loc, &max_loc);
   cout << "cpu min: " << min << ", max: " << max << endl;
   cout << "cpu min location: " << min_loc.x << "," << min_loc.y
           << ", max location: "  << max_loc.x << "," << max_loc.y  << endl;

OUTPUT:
gpu min: -0.000928892, max: 4.89271
gpu min location: 0,0, max location: 2,4

cpu min: 0, max: 0.0919901
cpu min location: 0,0, max location: 4,0